PURPOSE: To obtain the segmenting method and phoneme classifying method which provide a system with generally superior extendability and facilitates speaker adaption.
CONSTITUTION: Speech feature vectors extracted by speech analysis parts 21 and 22 are converted by a projection vector arithmetic part 3 and a projection operator memory 4 into dimension-reduced projection vectors and segmentation or phoneme classification is carried out according to a time series of the projection vectors. A projection operator stored in the memory 4 is generated by analyzing principal components of phoneme learnt samples and the projection vectors carry the majority of speech features on a low level and are artificially fixed as important elements carrying sound source features and glottis features. Consequently, the extendability of the system is rich and the speaker adaption is easily performed when the projection operators is changed.
ODA KEISUKE
WATANABE AKIHIKO
FUKAZAWA ATSUSHI